Rhododendron latoucheae Franch. (R. latoucheae) is a valuable woody plant known for its high ornamental value. While purple flowers are a distinct and attractive variant phenotype of R. latoucheae, the underlying mechanism regulating its flower color is still poorly understood. To investigate the molecular regulatory mechanism responsible for the variation in flower color, we selected plants with white-pink and purple petals as the object and conducted analyses of metabolites, key genes, and transcription factors associated with flower color. A combined metabolome–transcriptome analysis was performed, and the expression of key genes was subsequently verified through qRT-PCR experiments. The results of our study demonstrated a significant enrichment of differential metabolites in the flavonoid metabolic pathway. Changes in anthocyanin content followed the same trend as the observed flower color variations, specifically showing significant correlations with the contents of malvidin-3-O-glucoside, dihydromyricetin, gallocatechin, and peonidin-3-O-glucoside. Furthermore, we identified three key structural genes (F3GT1, LAR, ANR) and four transcription factors (bHLH130, bHLH41, bHLH123, MYB4) that are potentially associated with the biosynthesis of flavonoid compounds, thereby influencing the appearance of purple flower color in R. latoucheae.

Wildfire regimes in mountain forest landscapes are shaped by complex interactions among topography, climate, vegetation, and human activity. However, predicting and interpreting fire occurrence in topographically heterogeneous regions remains challenging because fire–environment relationships vary strongly across elevation gradients and temporal scales. This study developed a hybrid machine-learning framework integrating an Information Value Model (IVM), Random Forest (RF), and Convolutional Neural Network (CNN) to decode elevation-mediated wildfire regimes in western Sichuan, China, a mountainous forest region characterized by strong vertical environmental gradients and high ecological conservation value. Multi-source datasets, including Moderate Resolution Imaging Spectroradiometer (MODIS) burned-area records, topographic variables, monthly meteorological data, vegetation indices, land-cover information, and human-accessibility proxies, were integrated at a 500 m spatial resolution. Environmentally comparable non-fire samples were generated from unburned vegetated pixels, and model training, RF-based feature selection, hyperparameter tuning using Particle Swarm Optimization (PSO), and performance evaluation were conducted within a nested spatial block cross-validation framework. The model produced continuous wildfire occurrence probabilities and showed strong discriminatory performance under the adopted validation protocol, with AUC values exceeding 0.95 across temporal datasets and low probability-error metrics. RF importance and correlation analyses identified mean temperature, elevation, and precipitation as the dominant predictors of wildfire probability. Spatial analyses revealed pronounced elevation-mediated differentiation in wildfire regimes: low-elevation valleys showed higher fire probability and stronger associations with human-accessibility proxies, whereas high-elevation plateau areas exhibited lower and more scattered fire patterns associated with climatic constraints. Seasonal and monthly analyses further showed that winter and spring fires dominated the regional fire regime, with risk intensifying during the pre-monsoon dry period. By combining probabilistic fire-risk mapping, spatial-context learning, and elevation-gradient interpretation, this study provides a transferable framework for understanding wildfire regimes in complex mountain forest landscapes. The findings support adaptive forest fire management, targeted monitoring, and risk zoning in mountainous regions where forest ecosystems, human activities, and conservation values intersect.